U.S. patent application number 16/738926 was filed with the patent office on 2020-07-16 for apparatus and method for generating 3-dimensional full body skeleton model using deep learning.
This patent application is currently assigned to Electronics and Telecommunications Research Institute. The applicant listed for this patent is Electronics and Telecommunications Research Institute. Invention is credited to Hang-Kee KIM, Ki-Hong KIM, Ki-Suk LEE.
Application Number | 20200226827 16/738926 |
Document ID | 20200226827 / US20200226827 |
Family ID | 71516784 |
Filed Date | 2020-07-16 |
Patent Application | download [pdf] |
![](/patent/app/20200226827/US20200226827A1-20200716-D00000.png)
![](/patent/app/20200226827/US20200226827A1-20200716-D00001.png)
![](/patent/app/20200226827/US20200226827A1-20200716-D00002.png)
![](/patent/app/20200226827/US20200226827A1-20200716-D00003.png)
![](/patent/app/20200226827/US20200226827A1-20200716-D00004.png)
![](/patent/app/20200226827/US20200226827A1-20200716-D00005.png)
![](/patent/app/20200226827/US20200226827A1-20200716-D00006.png)
United States Patent
Application |
20200226827 |
Kind Code |
A1 |
KIM; Hang-Kee ; et
al. |
July 16, 2020 |
APPARATUS AND METHOD FOR GENERATING 3-DIMENSIONAL FULL BODY
SKELETON MODEL USING DEEP LEARNING
Abstract
Disclosed herein are an apparatus and method for generating a
skeleton model using deep learning. The method for generating a 3D
full-body skeleton model using deep learning, performed by the
apparatus for generating the 3D full-body skeleton model using deep
learning, includes generating training data using deep learning by
receiving a 2D X-ray image for training, analyzing the 2D X-ray
image of a user using the training data, and generating a 3D
full-body skeleton model by registering 3D local part bone models
generated from the result of analyzing the 2D X-ray image of the
user.
Inventors: |
KIM; Hang-Kee; (Daejeon,
KR) ; KIM; Ki-Hong; (Sejong-si, KR) ; LEE;
Ki-Suk; (Daejeon, KR) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Electronics and Telecommunications Research Institute |
Daejeon |
|
KR |
|
|
Assignee: |
Electronics and Telecommunications
Research Institute
Daejeon
KR
|
Family ID: |
71516784 |
Appl. No.: |
16/738926 |
Filed: |
January 9, 2020 |
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06T 7/337 20170101;
G06N 7/005 20130101; G06T 17/10 20130101; G06N 20/00 20190101; G06T
2207/20081 20130101; G06T 2207/20076 20130101; G06T 2207/10116
20130101; G06T 7/0012 20130101 |
International
Class: |
G06T 17/10 20060101
G06T017/10; G06T 7/00 20060101 G06T007/00; G06T 7/33 20060101
G06T007/33; G06N 7/00 20060101 G06N007/00; G06N 20/00 20060101
G06N020/00 |
Foreign Application Data
Date |
Code |
Application Number |
Jan 10, 2019 |
KR |
10-2019-0003537 |
Claims
1. An apparatus for generating a 3D full-body skeleton model using
deep learning, comprising: one or more processors; memory; and one
or more programs, wherein: the one or more programs are stored in
the memory and executed by the one or more processors, and the one
or more processors execute the one or more programs so as to
generate training data using deep learning by receiving a 2D X-ray
image for training, to analyze a 2D X-ray image of a user using the
training data, and to generate the 3D full-body skeleton model by
registering a 3D local part bone model generated from a result of
analyzing the 2D X-ray image of the user.
2. The apparatus of claim 1, wherein the one or more processors
generate the training data by extracting a feature point and a
boundary from the 2D X-ray image for training and by learning the
extracted feature point and boundary using deep learning.
3. The apparatus of claim 2, wherein the one or more processors are
configured to: set an initial feature point in order to recognize
the feature point in the 2D X-ray image for training, specify a
preset area within a preset distance from the initial feature
point, and learn a set within the preset area as the feature
point.
4. The apparatus of claim 3, wherein the one or more processors
generate the training data using a radiographic image captured
using at least one of CT and MRI in addition to the 2D X-ray image
for training.
5. The apparatus of claim 4, wherein the one or more processors
change a parameter of the radiographic image using a statistical
shape model.
6. The apparatus of claim 1, wherein the 2D X-ray image of the user
is acquired in such a way that an X-ray of a predefined body part,
among body parts of the user, in at least one posture is taken from
at least one direction.
7. The apparatus of claim 6, wherein the one or more processors
extract a feature point and a boundary from the X-ray image of the
user using the training data and determine the body part of the
user, the direction from which the X-ray is taken, and the posture
of the body part based on the feature point and the boundary,
thereby generating the 3D local part bone model.
8. The apparatus of claim 1, wherein the one or more processors
calculate a parameter for minimizing a difference value caused by
transforming a feature point and a boundary of the 3D local part
bone model into a feature point and a boundary of a statistical
shape model corresponding thereto.
9. The apparatus of claim 8, wherein the one or more processors
place the 3D local part bone models at locations on a 3D coordinate
system corresponding to body parts of the user and transform a
connection part between the 3D local part bone models using the
statistical shape model, thereby generating the 3D full-body
skeleton model.
10. The apparatus of claim 9, wherein the one or more processors
calculate a connection part parameter for minimizing a difference
value between a shape of the connection part and a shape
transformed from the connection part using the statistical shape
model in order to connect the 3D local part bone models with each
other.
11. A method for generating a 3D full-body skeleton model using
deep learning, performed by an apparatus for generating the 3D
full-body skeleton model using deep learning, comprising:
generating training data using deep learning by receiving a 2D
X-ray image for training; analyzing a 2D X-ray image of a user
using the training data; and generating the 3D full-body skeleton
model by registering a 3D local part bone model generated from a
result of analyzing the 2D X-ray image of the user.
12. The method of claim 11, wherein generating the training data is
configured to generate the training data by extracting a feature
point and a boundary from the 2D X-ray image for training and by
learning the extracted feature point and boundary using deep
learning.
13. The method of claim 12, wherein generating the training data is
configured to: set an initial feature point in order to recognize
the feature point in the 2D X-ray image for training, specify a
preset area within a preset distance from the initial feature
point, and learn a set within the preset area as the feature
point.
14. The method of claim 13, wherein generating the training data is
configured to generate the training data using a radiographic image
captured using at least one of CT and MRI in addition to the 2D
X-ray image for training.
15. The method of claim 14, wherein generating the training data is
configured to change a parameter of the radiographic image using a
statistical shape model.
16. The method of claim 11, wherein the 2D X-ray image of the user
is acquired in such a way that an X-ray of a predefined body part,
among body parts of the user, in at least one posture is taken from
at least one direction.
17. The method of claim 16, wherein analyzing the 2D X-ray image of
the user is configured to extract a feature point and a boundary
from the X-ray image of the user using the training data and to
determine the body part of the user, the direction from which the
X-ray is taken, and the posture of the body part based on the
feature point and the boundary, thereby generating the 3D local
part bone model.
18. The method of claim 17, wherein registering the 3D local part
bone model is configured to calculate a parameter for minimizing a
difference value caused by transforming a feature point and a
boundary of the 3D local part bone model into a feature point and a
boundary of a statistical shape model corresponding thereto.
19. The method of claim 18, wherein registering the 3D local part
bone model is configured to place the 3D local part bone models at
locations on a 3D coordinate system, corresponding to body parts of
the user, and to transform a connection part between the 3D local
part bone models using the statistical shape model, thereby
generating the 3D full-body skeleton model.
20. The method of claim 19, wherein registering the 3D local part
bone model is configured to calculate a connection part parameter
for minimizing a difference value between a shape of the connection
part and a shape transformed from the connection part using the
statistical shape model in order to connect the 3D local part bone
models with each other.
Description
CROSS REFERENCE TO RELATED APPLICATION
[0001] This application claims the benefit of Korean Patent
Application No. 10-2019-0003537, filed Jan. 10, 2019, which is
hereby incorporated by reference in its entirety into this
application.
BACKGROUND OF THE INVENTION
1. Technical Field
[0002] The present invention relates generally to deep-learning
technology and three-dimensional (3D) model construction
technology, and more particularly to technology for generating a
skeleton model using deep learning.
2. Description of the Related Art
[0003] When disease is diagnosed by analyzing the shape of a
skeleton, construction of a 3D model of the full-body skeleton of a
user enables more accurate diagnosis compared to when diagnosis is
made based on a 2D image of a localized body part. Also, analysis
of the 3D full-body skeleton model of a user may improve the
precision of an advance treatment plan for treating disease, and
the possibility of future disease may be more accurately predicted
through physical simulation of the corresponding skeleton
model.
[0004] As methods for configuring a 3D full-body skeleton model
representing a human body, there are multiple methods using any of
various medical devices, such as a CT/MRI device, an X-Ray device,
an external-body scan device, a body composition measurement device
(in-body measurement equipment), and the like.
[0005] Here, the most accurate and reliable method is configuring a
3D skeleton model based on data acquired through Computed
Tomography/Magnetic Resonance Imaging (CT/MRI). However, this
method incurs higher expenses for acquiring data than when other
devices are used, and it takes a lot of time to prepare a capture
device and to capture an image. Further, because CT causes exposure
to a large amount of radiation, only a body part, the image of
which is required for disease diagnosis, is captured, rather than
capturing a full body. Also, MRI is superior for extracting
features of organs, but produces less accurate data for skeletal
parts.
[0006] When an external-body scan device is used, the device is
inexpensive, there is no problem of radiation exposure, and it
takes a short time to capture an image. However, there is a
limitation in that a user must wear tight-fitting clothes in order
to accurately measure the body. In spite of this limitation, 3D
external body data may be constructed, but this method is of
limited usefulness as a method for constructing a 3D skeleton model
based on bones inside a body, which results in low accuracy.
Alternatively, it is possible to use a method of predicting an
internal skeleton from an external body shape using a statistical
scheme, but an error may be introduced during prediction.
[0007] As the simplest method, there is a method of predicting a 3D
skeletal structure inside a body by analyzing body composition, but
the accuracy thereof is lower than the method of predicting a
skeleton using an external-body scan device, and thus this method
is rarely used.
[0008] Due to the above-described problems, it is difficult to use
the existing methods, such as CT/MRI or the like, in order to
construct a 3D full-body model. When capturing and 3D-modeling
processes are periodically performed in order to monitor a change
in the 3D model of a user, the above-described problems, such as
radiation exposure, high expense, and the like, are made worse,
which lowers the usefulness of the method.
[0009] Meanwhile, Korean Patent No. 10-1921988, titled "Method for
creating personalized 3D skeleton model", discloses a method for
creating a 3D skeleton model of a patient by analyzing data on
respective bones corresponding to specific body parts of a user
with reference to a statistical model.
SUMMARY OF THE INVENTION
[0010] An object of the present invention is to save the expense of
constructing a 3D full-body skeleton model and to raise the
accuracy of prediction of a skeleton.
[0011] Another object of the present invention is to improve the
accuracy of disease diagnosis using a 3D full-body skeleton model
and to improve the precision of an advance treatment plan for
treating disease.
[0012] A further object of the present invention is to raise the
accuracy of prediction of the possibility of future disease through
physical simulation of a 3D full-body skeleton model.
[0013] In order to accomplish the above objects, an apparatus for
generating a 3D full-body skeleton model using deep learning
according to an embodiment of the present invention includes one or
more processors, memory, and one or more programs. The one or more
programs may be stored in the memory and executed by the one or
more processors, and the one or more processors may execute the one
or more programs so as to generate training data using deep
learning by receiving a 2D X-ray image for training, to analyze a
2D X-ray image of a user using the training data, and to generate
the 3D full-body skeleton model by registering (matching) a 3D
local part bone model generated from the result of analyzing the 2D
X-ray image of the user.
[0014] Here, the one or more processors may generate the training
data by extracting a feature point and a boundary from the 2D X-ray
image for training and by learning the extracted feature point and
boundary using deep learning.
[0015] Here, the one or more processors may set an initial feature
point in order to recognize the feature point in the 2D X-ray image
for training, specify a preset area within a preset distance from
the initial feature point, and learn a set within the preset area
as the feature point.
[0016] Here, the one or more processors may generate the training
data using a radiographic image captured using at least one of CT
and MRI in addition to the 2D X-ray image for training.
[0017] Here, the one or more processors may change a parameter of
the radiographic image using a statistical shape model.
[0018] Here, the 2D X-ray image of the user may be acquired in such
a way that an X-ray of a predefined body part, among body parts of
the user, in at least one posture is taken from at least one
direction.
[0019] Here, the one or more processors may extract a feature point
and a boundary from the X-ray image of the user using the training
data and determine the body part of the user, the direction from
which the X-ray is taken, and the posture of the body part based on
the feature point and the boundary, thereby generating the 3D local
part bone model.
[0020] Here, the one or more processors may calculate a parameter
for minimizing a difference value caused by transforming the
feature point and the boundary of the 3D local part bone model into
the feature point and the boundary of a statistical shape model
corresponding thereto.
[0021] Here, the one or more processors may place the 3D local part
bone models at locations on a 3D coordinate system corresponding to
body parts of the user and transform a connection part between the
3D local part bone models using the statistical shape model,
thereby generating the 3D full-body skeleton model.
[0022] Here, the one or more processors may calculate a connection
part parameter for minimizing a difference value between the shape
of the connection part and a shape transformed from the connection
part using the statistical shape model in order to connect the 3D
local part bone models with each other.
[0023] Also, in order to accomplish the above objects, a method for
generating a 3D full-body skeleton model using deep learning,
performed by an apparatus for generating the 3D full-body skeleton
model using deep learning, according to an embodiment of the
present invention includes generating training data using deep
learning by receiving a 2D X-ray image for training, analyzing a 2D
X-ray image of a user using the training data, and generating the
3D full-body skeleton model by registering (matching) a 3D local
part bone model generated from the result of analyzing the 2D X-ray
image of the user.
[0024] Here, generating the training data may be configured to
generate the training data by extracting a feature point and a
boundary from the 2D X-ray image for training and by learning the
extracted feature point and boundary using deep learning.
[0025] Here, generating the training data may be configured to set
an initial feature point in order to recognize the feature point in
the 2D X-ray image for training, to specify a preset area within a
preset distance from the initial feature point, and to learn a set
within the preset area as the feature point.
[0026] Here, generating the training data may be configured to
generate the training data using a radiographic image captured
using at least one of CT and MRI in addition to the 2D X-ray image
for training.
[0027] Here, generating the training data may be configured to
change a parameter of the radiographic image using a statistical
shape model.
[0028] Here, the 2D X-ray image of the user may be acquired in such
a way that an X-ray of a predefined body part, among body parts of
the user, in at least one posture is taken from at least one
direction.
[0029] Here, analyzing the 2D X-ray image of the user may be
configured to extract a feature point and a boundary from the X-ray
image of the user using the training data and to determine the body
part of the user, the direction from which the X-ray is taken, and
the posture of the body part based on the feature point and the
boundary, thereby generating the 3D local part bone model.
[0030] Here, registering (matching) the 3D local part bone model
may be configured to calculate a parameter for minimizing a
difference value caused by transforming the feature point and the
boundary of the 3D local part bone model into the feature point and
the boundary of a statistical shape model corresponding
thereto.
[0031] Here, registering (matching) the 3D local part bone model
may be configured to place the 3D local part bone models at
locations on a 3D coordinate system, corresponding to body parts of
the user, and to transform a connection part between the 3D local
part bone models using the statistical shape model, thereby
generating the 3D full-body skeleton model.
[0032] Here, registering (matching) the 3D local part bone model
may be configured to calculate a connection part parameter for
minimizing a difference value between the shape of the connection
part and a shape transformed from the connection part using the
statistical shape model in order to connect the 3D local part bone
models with each other.
BRIEF DESCRIPTION OF THE DRAWINGS
[0033] The above and other objects, features and advantages of the
present invention will be more clearly understood from the
following detailed description taken in conjunction with the
accompanying drawings, in which:
[0034] FIG. 1 is a block diagram that shows an apparatus for
generating a 3D full-body skeleton model using deep learning
according to an embodiment of the present invention;
[0035] FIG. 2 is a block diagram that specifically shows an example
of the training-data generation unit illustrated in FIG. 1;
[0036] FIG. 3 is a view that shows data that is necessary in order
to generate training data according to an embodiment of the present
invention;
[0037] FIG. 4 is a view that shows an example of the full-body
skeleton model generation unit illustrated in FIG. 1;
[0038] FIG. 5 is a flowchart that shows a method for generating a
3D full-body skeleton model using deep learning according to an
embodiment of the present invention;
[0039] FIG. 6 is a flowchart that specifically shows an example of
the step of generating training data illustrated in FIG. 5;
[0040] FIG. 7 is a flowchart that specifically shows an example of
the step of generating the full-body skeleton model illustrated in
FIG. 5; and
[0041] FIG. 8 is a view that shows a computer system according to
an embodiment of the present invention.
DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0042] The present invention will be described in detail below with
reference to the accompanying drawings. Repeated descriptions and
descriptions of known functions and configurations which have been
deemed to unnecessarily obscure the gist of the present invention
will be omitted below. The embodiments of the present invention are
intended to fully describe the present invention to a person having
ordinary knowledge in the art to which the present invention
pertains. Accordingly, the shapes, sizes, etc. of components in the
drawings may be exaggerated in order to make the description
clearer.
[0043] Throughout this specification, the terms "comprises" and/or
"comprising" and "includes" and/or "including" specify the presence
of stated elements but do not preclude the presence or addition of
one or more other elements unless otherwise specified.
[0044] An apparatus and method for generating a 3D full-body
skeleton model using deep learning according to an embodiment of
the present invention may enable a 3D full-body skeleton model to
be generated using data acquired from an X-ray device. X-ray data
has lower accuracy than a CT/MRI image, but may be acquired at low
cost. Also, when an X-ray is taken, the amount of radiation
exposure is lower than when a CT scan is performed, and X-ray data
is advantageous in extracting skeletal data compared to MRI. Here,
a skeleton may be predicted using several X-ray images of bones,
rather than using an external-body scanner, whereby the accuracy
may be raised compared to when prediction is performed using an
external-body scanner.
[0045] The present invention may generate 3D local part bone models
(of, for example, a pelvis, a spine, a femur, a fibula, a thorax,
and the like) using X-ray data, and may generate a full-body
skeleton model using the local part bone models.
[0046] Hereinafter, a preferred embodiment of the present invention
will be described in detail with reference to the accompanying
drawings.
[0047] FIG. 1 is a block diagram that shows an apparatus for
generating a 3D full-body skeleton model using deep learning
according to an embodiment of the present invention. FIG. 2 is a
block diagram that specifically shows an example of the
training-data generation unit illustrated in FIG. 1. FIG. 3 is a
view that shows data required for generating training data
according to an embodiment of the present invention. FIG. 4 is a
view that specifically shows an example of the full-body skeleton
model generation unit illustrated in FIG. 1.
[0048] Referring to FIG. 1, the apparatus for generating a 3D
full-body skeleton model using deep learning according to an
embodiment of the present invention includes a training-data
generation unit 110, a local part bone model registration unit 120,
and a full-body skeleton model generation unit 130.
[0049] The training-data generation unit 110 may generate training
data using deep learning by receiving a 2D X-ray image for
training.
[0050] Here, the training-data generation unit 110 may extract a
feature point and a boundary from the 2D X-ray image for training,
and may generate training data by learning the extracted feature
point and boundary using deep learning.
[0051] Here, the training-data generation unit 110 sets an initial
feature point in order to recognize the feature point in the 2D
X-ray image for training, and specifies a preset area within a
preset distance from the initial feature point, thereby learning a
set in the preset area as the feature point.
[0052] For example, the training-data generation unit 110 may
recognize an image and extract a feature point and a boundary by
employing a method of processing a bounding box for representing
the area of a recognized object (e.g., Yolo, RetinaNet, SSD, or the
like).
[0053] Here, the training-data generation unit 110 may set a
bounding box including a peripheral area in order to learn the
location of the feature point, and may perform training by
regarding a specified set as training data. Because the feature
point is given in a point format, the center of the bounding box
may be set as the feature point, or the four edges thereof may be
set as the feature points.
[0054] Here, the training-data generation unit 110 may set multiple
bounding boxes and use a combination thereof (For example, when
four neighboring bounding boxes are present, the locations of
feature points are set to point the edges at which the four boxes
are close to each other, and training with respect to the four
feature points may be performed individually. When they are
recognized, the average of the four edges of the recognized four
boxes may be regarded as the feature points.)
[0055] Here, the training-data generation unit 110 may extract a
boundary using any of recent deep-learning segmentation techniques
(e.g., Mask RCNN, semantic segmentation, DeepLab, Polygon-RNN, and
the like).
[0056] Here, the training-data generation unit 110 may set the area
to be recognized using a boundary, and may learn a set of
boundaries by regarding the same as training data. Here, the
boundary may be learned using a hierarchical structure. (For
example, after the entire femur is recognized using a single
boundary, training is performed such that a femoral head area is
recognized as a sub-boundary area, whereby a recognition range may
be scaled down in phases.)
[0057] Here, the training-data generation unit 110 may construct a
2D X-ray data set.
[0058] Here, the training-data generation unit 110 may perform
training using a set of training data (X-ray images, annotation,
and the like), and may generate training data (a weight and the
like) as the result of training.
[0059] First, full-body biplanar X-ray images (e.g., EOS imaging or
the like) may be the first candidate of the 2D data set. The
corresponding data is acquired by capturing a full-body image, and
because prearranged image data perpendicularly projected from
frontal and lateral views is acquired, information for constructing
a 3D full-body model may be easily acquired and analyzed. However,
because a device capable of capturing full-body biplanar X-ray
images is expensive and takes up a lot of space, it is difficult to
equip general clinics with such a device. Accordingly, there are
few images acquired using the biplanar X-ray data, and there is not
enough published data. Therefore, x-ray data acquired in such a way
that images of various body parts, such as a chest, knees, a pelvis
and the like, in various postures, such as bending, stretching, and
the like, are captured from different directions, such as an
anteroposterior or posteroanterior view (PA, AP) or lateral view,
may be used in order to make diagnosis of disease.
[0060] The present invention uses the above-mentioned various kinds
of data as training data. In the state in which the type and
location of the feature points to be used for training of each body
part are predefined or in which the shape of the boundary to be
used for training of each body part is predefined, when the
predefined part is found in an X-ray image for individual training,
the part may be marked according to predefined content. (Here, it
is possible to mark the predefined part in advance, but marking
using a semi-automatic method, in which the feature point or the
boundary is automatically extracted from the corresponding
individual training data (an X-ray image or the like) using data (a
weight or the like) on which training has been performed and then a
portion having an error is adjusted, may be employed in order to
reduce effort.)
[0061] Here, because the task of generating training data by
setting a feature point or a boundary area in an X-ray image
requires medical knowledge and because many errors may be caused
when ordinary people arbitrarily set the feature point or the
boundary area, the training-data generation unit 110 may need
reference data generated by experts.
[0062] The present invention may take such data as a candidate for
training data. However, in many cases, a desired form of feature
point or boundary is not marked, and thus an additional task may be
required. As described above, published data may be used as
training data, or X-ray data of a user may be used as training data
after obtaining the user's consent, in which case the data should
be strictly managed to prevent leakage of private information.
[0063] Here, the training-data generation unit 110 may generate
training data using a radiographic image acquired using at least
one of CT and MRI in addition to the 2D X-ray images for
training.
[0064] Referring to FIG. 3, in order to overcome the lack of 2D
X-ray data for training, the present invention may additionally use
Digitally Reconstructed Radiograph (DRR) data corresponding to a
digitally reconstructed radiographic image.
[0065] Here, when the amount of original data for deep learning is
insufficient, the training-data generation unit 110 may use, along
with the original data, data having characteristics similar to
those of the original data as training data, thereby improving
performance. Here, DRR is regarded as such data having similar
characteristics.
[0066] Here, the training-data generation unit 110 may generate a
pseudo X-ray image based on CT/MRI data corresponding to DRR or on
a 3D mesh model.
[0067] Here, the training-data generation unit 110 may generate DRR
data directly from CT/MRI data, or may derive a 3D model from
CT/MRI data and generate a DRR image through the 3D model.
[0068] Here, the training-data generation unit 110 may generate DRR
using different camera directions, or may generate DRR by
transforming a 3D model through a change in a parameter using a
statistical shape model (SSM), which will be described later. The
CT/MRI data or the 3D mesh model does not have to be full-body
data. Even though the CT/MRI data or the 3D mesh model pertains to
a local part, the training-data generation unit 110 may generate
DRR from X-ray data pertaining to the local part and use the same
as training data.
[0069] Also, the training-data generation unit 110 may change the
parameter of the radiographic image using a statistical shape
model.
[0070] Here, the training-data generation unit 110 may be required
to construct a statistical shape model (SSM) for generating a 3D
skeleton model in addition to being required to generate training
data (a weight or the like) using 2D X-ray data for training.
[0071] Here, the training-data generation unit 110 uses the
statistical shape model of a skeleton in order to perform
statistical analysis based on 3D skeleton data, and changes an
average-shaped 3D model, which is derived from the analysis result,
by adjusting a PCA parameter, thereby representing the shapes of
skeletons of different users.
[0072] Here, for the statistical shape model, the training-data
generation unit 110 may generate various forms of 3D skeleton model
data for each part in order to perform statistical analysis.
[0073] Here, the training-data generation unit 110 may construct a
3D model set through full-body CT/MRI data.
[0074] Here, the training-data generation unit 110 may generate a
statistical shape model for each part (e.g., a femur, a pelvis, a
spine, or the like) of the body of a user.
[0075] Here, the training-data generation unit 110 may need at
least one 3D full-body skeleton model.
[0076] Here, when it later constructs a 3D skeleton model of a
user, the training-data generation unit 110 may use a previously
constructed 3D full-body skeleton model if a 3D model was
constructed only for a local part because there is no full-body
X-ray image of the user. (Here, if two or more previously
constructed 3D full-body skeleton models are present, a statistical
shape model (SSM) of the corresponding skeleton model is
constructed in advance).
[0077] Referring to FIG. 2, the training-data generation unit 110
may include a feature-point-setting unit 111, a
feature-point-learning unit 112, and a data-learning unit 113.
[0078] The feature-point-setting unit 111 may set an initial
feature point in order to recognize the feature point in the 2D
X-ray image for training and specify a preset area within a preset
distance from the initial feature point, and the
feature-point-learning unit 112 may learn a set in the preset area
as the feature point.
[0079] Here, the feature-point-setting unit 111 may extract a
feature point and a boundary from the 2D X-ray image for training,
the feature-point-learning unit 112 may learn the extracted feature
point and boundary using deep learning, and the data-learning unit
113 may generate training data.
[0080] The data-learning unit 113 may generate training data using
a radiographic image captured using at least one of CT and MRI in
addition to the 2D X-ray images for training.
[0081] Here, the data-learning unit 113 may change the parameter of
the radiographic image using a statistical shape model.
[0082] The training data may include data on which training has
been performed through deep learning for recognition and
segmentation of each part in a 2D X-ray image, a 2D data set for
learning the data, statistical shape model (SSM) data acquired
through statistical model analysis of a 3D skeleton model, a 3D
data set for analysis, and the like.
[0083] Also, the training-data generation unit 110 may perform
segmentation in order to extract the feature point of a desired
bone (e.g., a greater trochanter, a lesser trochanter, a condyle,
the edge of an inner condyle, the center point of a femoral head,
or the like in a femur) or the boundary of a shape (e.g., the
entire boundary of a femur) through deep learning when a 3D
skeleton model is constructed using the X-ray image of a user, and
may calculate trained data (a deep-learning weight or the like) in
advance.
[0084] The local part bone model registration unit 120 may analyze
the 2D X-ray image of a user using the training data, and may
register a 3D local part bone model, which is generated from the
result of analysis of the 2D X-ray image of the user.
[0085] Here, the 2D X-ray image of the user may be acquired in such
a way that an X-ray of a predefined body part, among the body parts
of the user, in at least one posture is taken from at least one
direction.
[0086] Here, the local part bone model registration unit 120 may
extract a feature point and a boundary from the X-ray image of the
user using the training data and determine the body part of the
user, and the direction and posture in which the X-ray of the body
part is taken based on the feature point and the boundary, thereby
generating a 3D local part model.
[0087] Here, the local part bone model registration unit 120 may
extract the feature point or the boundary from the X-ray image of
the user.
[0088] Here, the local part bone model registration unit 120 may
perform calculation through deep-learning technology using trained
data (a weight or the like).
[0089] Here, when the X-ray image does not correspond to full-body
data, the local part bone model registration unit 120 may derive
the feature point and the boundary of only a local part calculable
from the corresponding data, and may calculate the body part and
the posture corresponding to the X-ray image.
[0090] Here, when the device used for capturing the X-ray image of
the user is not a previously calculated vertical biplanar X-ray
device, if data acquired by capturing the same part in different
directions is present in two or more X-ray images, the local part
bone model registration unit 120 extracts a feature point or a
boundary therefrom. Here, when the same feature point set (e.g., a
greater trochanter of a left leg) is derived from the two or more
different X-ray images, the local part bone model registration unit
120 may derive the relative location at which the X-ray is taken
and the direction in which the X-ray is taken using the
corresponding points (through a method such as solvePnP of OpenCV
or the like).
[0091] Here, when the relative location and orientation of the
camera (the device for taking an X-ray) are derived as described
above, the 3D locations of the feature points (or boundary) of a
corresponding pair are also derived, and the local part bone model
registration unit 120 may represent the remaining feature points,
which are not common to the different X-ray images, using a single
coordinate system.
[0092] Here, because the remaining points are not represented in a
3D coordinate system, the local part bone model registration unit
120 may use a 2D coordinate value to which a transformation value
based on the relative location and orientation of the device is
applied on the plane of a 2D coordinate system.
[0093] Here, the local part bone model registration unit 120 may
calculate a parameter for minimizing a difference value caused by
transforming the feature point and the boundary of the 3D local
part bone model into the feature point and the boundary of the
statistical shape model corresponding thereto.
[0094] For example, the local part bone model registration unit 120
may calculate the parameter of the optimum statistical shape model
(SSM) that best matches the extracted feature point or boundary of
each part.
[0095] Here, the local part bone model registration unit 120 may
use a set of PCA parameters, which is used in the SSM, along with a
transformation value for translation, rotation, and scaling of the
SSM, as the parameter required for optimization (in which case the
number of PCA parameters is limited to multiples of ten rather than
using all of the parameters, and as the number of parameters
increases, it is possible to respond to more diverse variation, but
more time is spent deriving the optimum value).
[0096] Here, the local part bone model registration unit 120 may
use various methods in order to optimize the parameter, such as a
least-squares method, a Gauss-Newton method, or a
Levenberg-Marquardt optimization method.
[0097] Here, when optimization is performed, it is necessary to
calculate a cost value, and the local part bone model registration
unit 120 may use the difference between the location of the feature
point derived from the X-ray image and the location of the feature
point corresponding thereto, which is derived from the orthogonal
projection of a shape model, which is acquired through a change in
the parameter of the SSM, onto a 2D coordinate system, when
optimization using the feature point is performed.
[0098] Here, when the boundary is used for optimization, the local
part bone model registration unit 120 compares the boundary derived
from the X-ray image with the boundary of the orthogonal projection
of a shape model, which is acquired through a change in the
parameter of the SSM, onto the 2D coordinate system and calculates
the difference therebetween as the cost value. Here, parameter
optimization may be performed such that the difference is
minimized.
[0099] The full-body skeleton model generation unit 130 may
generate a 3D full-body skeleton model by transforming the result
of registration of the 3D local part bone model.
[0100] Here, the full-body skeleton model generation unit 130
places the 3D local part bone models at the locations on the 3D
coordinate system corresponding to the body parts of the user, and
transforms a connection part between the 3D local part bone models
using the statistical shape model, thereby generating a 3D
full-body skeleton model.
[0101] Here, in order to connect the 3D local part bone models with
each other, the full-body skeleton model generation unit 130 may
calculate a connection part parameter for minimizing the difference
value between the shape of the connection part and the shape
transformed from the connection part using the statistical shape
model.
[0102] For example, when the 3D skeleton model is not derived from
the initial full-body X-ray image, the full-body skeleton model
generation unit 130 may acquire only the 3D local part bone model
derived from the X-ray image of the captured body part. In this
case, the height and weight of the user are acquired, and the
external body data and the length of a joint may be calculated
using an external-body scan device.
[0103] Here, the full-body skeleton model generation unit 130 may
acquire the length of the joint using external body scan data
(e.g., Kinect, Open Pose, or the like).
[0104] Here, the full-body skeleton model generation unit 130 may
calculate a model that matches the characteristics of the user by
transforming a previously constructed full-body model based on the
above-described information.
[0105] Here, the full-body skeleton model generation unit 130 may
transform the previously constructed full-body model through
optimization of a transformation value for translation, rotation,
and scaling, the PCA parameter used in the SSM of the 3D full-body
model, and additional parameters, such as the height or the like,
as in the case of calculation of the parameter.
[0106] Here, the full-body skeleton model generation unit 130 may
place the previously constructed 3D model of the local part in the
part corresponding thereto and calculate a transformation parameter
such that the 3D local part model is most smoothly connected.
[0107] Here, the full-body skeleton model generation unit 130 may
calculate the parameter derived from the difference value between
the transformed body part of the 3D full-body model and the shape
of the 3D body model of the corresponding local part as the cost
value for optimization.
[0108] Here, when the optimum full-body SSM is found, the full-body
skeleton model generation unit 130 places the 3D bone models of
respective body parts in the full-body model and deletes the parts
of the existing full-body model corresponding thereto such that the
3D bone models replace the deleted parts.
[0109] Here, when the replaced part is not smoothly connected, the
full-body skeleton model generation unit 130 may smoothly connect
the replaced part through blending and interpolation of a
difference in each vertex of the connection part.
[0110] Here, when it lacks the height, the weight, or some or all
of the external scan data, the full-body skeleton model generation
unit 130 may calculate the optimum transformation parameter of the
average 3D full-body model that matches the 3D local part model
without the corresponding information.
[0111] Referring to FIG. 4, the full-body skeleton model generation
unit 130 may include a model arrangement unit 131, a parameter
generation unit 132, and a matching-part-blending unit 133.
[0112] The model arrangement unit 131 places the 3D local part bone
models at the locations on the 3D coordinate system corresponding
to the body parts of the user and transforms a connection part
between the 3D local part bone models using the statistical shape
model, thereby generating a 3D full-body skeleton model.
[0113] The parameter generation unit 132 may calculate a connection
part parameter for minimizing the difference value between the
shape of the connection part and the shape that is transformed from
the connection part using the statistical shape model in order to
connect the 3D local part bone models with each other.
[0114] The matching-part-blending unit 133 may smoothly connect the
connection part of the 3D local part bone model through blending
and interpolation of a difference in each vertex of the connection
part when the connection part is not smoothly connected.
[0115] FIG. 5 is a flowchart that shows a method for generating a
3D full-body skeleton model using deep learning according to an
embodiment of the present invention. FIG. 6 is a flowchart that
specifically shows an example of the step of generating training
data illustrated in FIG. 5. FIG. 7 is a flowchart that specifically
shows an example of the step of generating a full-body skeleton
model illustrated in FIG. 5.
[0116] Referring to FIG. 5, in the method for generating a 3D
full-body skeleton model using deep learning according to an
embodiment of the present invention, first, training data may be
generated at step S210.
[0117] That is, at step S210, an initial feature point is set in
order to recognize a feature point in a 2D X-ray image for
training, a preset area within a preset distance from the initial
feature point is specified, and training is performed with respect
to a set in the preset area as the feature point.
[0118] Referring to FIG. 6, step S210 may be configured such that a
feature point and a boundary are extracted from the 2D X-ray image
for training at step S211, training is performed with respect to
the extracted feature point and boundary using deep learning at
step S212, and training data may be generated at step S213.
[0119] Here, at step S213, the training data may be generated using
a radiographic image captured using at least one of CT and MRI in
addition to the 2D X-ray image for training.
[0120] Here, at step S213, the parameter of the radiographic image
may be changed using a statistical shape model.
[0121] The training data may include data on which training has
been performed through deep learning for recognition and
segmentation of each part in the 2D X-ray image, a 2D data set for
learning the data, statistical shape model (SSM) data acquired
through statistical model analysis of a 3D skeleton model, a 3D
data set for analysis, and the like.
[0122] Also, at step S210, when the 3D skeleton model is
constructed using the X-ray image of a user, segmentation may be
performed in order to extract the feature point of a desired bone
(e.g., a greater trochanter, a lesser trochanter, a condyle, the
edge of an inner condyle, the center point of a femoral head, or
the like in a femur) or the boundary of a shape (e.g., the entire
boundary of a femur) through deep learning, and the trained data (a
deep-learning weight or the like) may be calculated in advance.
[0123] Also, at step S220, preprocessing data for a 2D medical
image and a 3D skeleton model may be generated.
[0124] At step S220, when the 3D skeleton model is constructed
using the X-ray image of a user, segmentation may be performed in
order to extract the feature point of a desired bone (e.g., a
greater trochanter, a lesser trochanter, a condyle, the edge of an
inner condyle, the center point of a femoral head, or the like in a
femur) or the boundary of a shape (e.g., the entire boundary of a
femur) through deep learning, and the trained data (a deep-learning
weight or the like) may be calculated in advance.
[0125] Also, the 2D X-ray image of the user may be analyzed using
the training data at step S230, and the 3D local part bone model,
generated from the result of analysis of the 2D X-ray image of the
user, may be registered (matched) at step S240.
[0126] Here, the 2D X-ray image of the user may be acquired in such
a way that an X-ray of a predefined body part, among the body parts
of the user, in at least one posture is taken from at least one
direction.
[0127] Here, at step S230, a feature point and a boundary are
extracted from the X-ray image of the user using the training data,
and the body part of the user and the direction and posture, in
which the X-ray of the body part is taken, are determined based on
the feature point and the boundary, whereby a 3D local part model
may be generated.
[0128] Here, at step S230, the feature point or the boundary may be
extracted from the X-ray image of the user.
[0129] Here, at step S230, calculation may be made based on a
deep-learning technique using trained data (a weight or the
like).
[0130] Here, at step S230, when the X-ray image does not correspond
to full-body data, the feature point and the boundary only of a
local part calculable from the corresponding data are extracted,
and the body part and the posture corresponding to the X-ray image
may be calculated.
[0131] Here, at step S230, if the device used for capturing the
X-ray image of the user is not a previously calculated vertical
biplanar X-ray device, when data acquired by capturing the same
part in different directions is present in two or more X-ray
images, a feature point or a boundary is derived therefrom. Here,
when the same feature point set (e.g., a greater trochanter of a
left leg) is derived from the two or more different X-ray images,
the relative location at which the X-ray is taken and the direction
in which the X-ray is taken may be derived using the corresponding
points (through a method such as solvePnP of OpenCV or the
like).
[0132] Here, at step S230, when the relative location and
orientation of the camera (the device for taking an X-ray) are
derived as described above, the 3D locations of the feature points
(or boundary) of a corresponding pair are derived. Further, the
remaining feature points, which are not common to the different
X-ray images, may be represented in a single coordinate system.
[0133] Here, at step S230, because the remaining points are not
represented in a 3D coordinate system, a 2D coordinate value to
which a transformation value based on the relative location and
orientation of the device is applied on the plane of a 2D
coordinate system may be used.
[0134] Here, at step S230, a parameter for minimizing the
difference value caused by transforming the feature point and the
boundary of the 3D local part bone model into those of the
statistical shape model corresponding thereto may be
calculated.
[0135] For example, at step S230, the parameter of the optimum
statistical shape model (SSM) that best matches the extracted
feature point or boundary of each part may be calculated.
[0136] Here, at step S230, as the parameter required for
optimization, a set of PCA parameters, which is used in the SSM,
may be used along with a transformation value for translation,
rotation, and scaling of the SSM (in which case the number of PCA
parameters is limited to multiples of ten rather than using all of
the parameters, and as the number of parameters increases, it is
possible to respond to more diverse variation, but more time is
spent deriving the optimum value).
[0137] Here, at step S230, in order to optimize the parameter, any
of various methods, such as a least-squares method, a Gauss-Newton
method, or a Levenberg-Marquardt optimization method, may be
used.
[0138] Here, at step S230, when optimization performed, it is
necessary to calculate a cost value. In the case of optimization
using a feature point, the difference between the location of the
feature point derived from the X-ray image and the location of the
feature point corresponding thereto, which is derived from the
orthogonal projection of a shape model acquired through a change in
the parameter of the SSM onto a 2D coordinate system, may be
used.
[0139] Here, at step S230, when a boundary is used for
optimization, the boundary derived from the X-ray image is compared
with the boundary of the orthogonal projection of a shape model,
which is acquired through a change in the parameter of the SSM,
onto the 2D coordinate system, and the difference therebetween is
calculated as a cost value. Here, parameter optimization may be
performed such that the difference is minimized.
[0140] Also, at step S240, the 3D local part model, generated using
the result of analysis of the X-ray image of the user, may be
registered (matched).
[0141] Also, at step S250, a full-body skeleton model may be
generated using the result of registration (matching) of the 3D
local part bone model.
[0142] Referring to FIG. 7, step S250 is configured such that the
3D local part bone models are placed at the locations on the 3D
coordinate system corresponding to the body parts of the user at
step S251, a parameter is generated at step S252, and a connection
part between the 3D local part bone models is transformed using the
statistical shape model and the matching parts are blended at step
S253, whereby a 3D full-body skeleton model may be generated at
step S254.
[0143] Here, at step S252, in order to connect the 3D local part
bone models with each other, a connection part parameter for
minimizing the difference value between the shape of the connection
part and the shape transformed from the connection part using the
statistical shape model may be calculated.
[0144] For example, at step S252, when a 3D skeleton model is not
derived from an initial full-body X-ray image, only 3D local part
bone models are derived from X-ray images acquired by capturing
respective body parts. In this case, the height and weight of the
user are acquired, and the external-body data and the lengths of
joints may be calculated using an external-body scan device.
[0145] Here, at step S252, the lengths of joints may be derived
using external-body scan data (e.g., Kinect, Open Pose or the
like).
[0146] Here, at step S252, a model that matches the characteristics
of the user may be calculated by transforming a previously
constructed full-body model based on the above-described
information.
[0147] Here, at step S252, transformation of the previously
constructed full-body model may be performed through optimization
of a transformation value for translation, rotation, and scaling,
the PCA parameter used in the SSM of the 3D full-body model, and
additional parameters, such as the height or the like, as in the
case of calculation of the parameter.
[0148] Here, at step S252, the previously constructed 3D model of
the local part is placed in the corresponding body part, and the
transformation parameter may be calculated such that the
corresponding local part model is most smoothly connected.
[0149] Here, at step S252, the parameter derived from the
difference value between the transformed body part of the 3D
full-body model and the shape of the 3D model of the corresponding
local part may be calculated as the cost value for
optimization.
[0150] Here, at step S252, when the optimum full-body SSM is found,
the 3D bone models for respective body parts are placed in the
full-body model, and the parts of the existing full-body model
corresponding thereto are deleted, whereby the 3D bone models
replace the deleted parts.
[0151] Here, at step S253, when the replaced part is not smoothly
connected, the replaced part may be smoothly connected through
blending and interpolation of a difference in each vertex of the
connection part.
[0152] Here, at step S252, when the height, the weight, or some or
all of external body scan data is insufficient, the optimum
transformation parameter of an average 3D full-body model matching
the 3D local part model may be calculated without the corresponding
information.
[0153] Here, at step S254, a 3D full-body skeleton model may be
finally generated.
[0154] FIG. 8 is a block diagram that shows a computer system
according to an embodiment of the present invention.
[0155] Referring to FIG. 8, the apparatus for generating a 3D
full-body skeleton model using deep learning according to an
embodiment of the present invention may be implemented in a
computer system 1100 including a computer-readable recording
medium. As shown in FIG. 8, the computer system 1100 may include
one or more processors 1110, memory 1130, a user-interface input
device 1140, a user-interface output device 1150, and storage 1160,
which communicate with each other via a bus 1120. Also, the
computer system 1100 may further include a network interface 1170
connected with a network 1180. The processor 1110 may be a central
processing unit or a semiconductor device for executing processing
instructions stored in the memory 1130 or the storage 1160. The
memory 1130 and the storage 1160 may be any of various types of
volatile or nonvolatile storage media. For example, the memory may
include ROM 1131 or RAM 1132.
[0156] Here, the apparatus for generating a 3D full-body skeleton
model using deep learning according to an embodiment of the present
invention includes one or more processors 1110, memory 1130, a
user-interface input device 1140, a user-interface output device
1150, and storage 1160, which communicate with each other via a bus
1120, and one or more programs. The one or more programs are stored
in the memory and executed by the one or more processors 1110. When
the one or more processors execute the one or more programs,
training data may be generated using deep learning by receiving a
2D X-ray image for training, the 2D X-ray image of a user may be
analyzed using the training data, and a 3D local part bone model
generated from the result of analysis of the 2D X-ray image of the
user may be registered.
[0157] Here, the one or more processors 1110 may perform the
functions of the training-data generation unit 110, the local part
bone model registration unit 120, and the full-body skeleton model
generation unit 130, described with reference to FIGS. 1 to 4, and
may operate based on the description made with reference to FIGS. 1
to 4, and thus a detailed description thereof will be omitted.
[0158] The present invention may save the expense of constructing a
3D full-body skeleton model, and may raise the accuracy of
prediction of a skeleton.
[0159] Also, the present invention may improve the accuracy of
disease diagnosis using a 3D full-body skeleton model, and may
improve the precision of an advance treatment plan for treating
disease.
[0160] Also, the present invention may raise the accuracy of
prediction of the possibility of future disease through physical
simulation of a 3D full-body skeleton model.
[0161] As described above, the apparatus and method for generating
a skeleton model using deep learning according to the present
invention are not limitedly applied to the configurations and
operations of the above-described embodiments, but all or some of
the embodiments may be selectively combined and configured, so that
the embodiments may be modified in various ways.
* * * * *